An explicitly-undef optional argument means absent, not bad

[revarbat]

BOSL2 reaches a builtin whose name it also defines by wrapping it in builtins.scad — a file that use <>s nothing, so the names inside it bind to the builtins rather than to BOSL2's own definitions:

// builtins.scad -- includes nothing, only ever use<>d
function _linear_solve(A, b) = linear_solve(A, b);

That wrapper has a fixed signature, so a determinant()-style caller that passes no right-hand side still forwards b as undef. We treated that as a bad argument — warning on every call and returning undef, which made determinant() unusable through the wrapper.

before:  linear_solve(A)         -> det = 5
         linear_solve(A, undef)  -> WARNING, det = undef      <-- what the wrapper produces
after:   both give det = 5, silently

Passing undef is idiomatically the same as not passing in OpenSCAD — it's how BOSL2 threads optional arguments through wrappers everywhere (_cylinder(h,r1,r2,center,r,d,d1,d2) forwards all eight, most of them undef). So undef now counts as absent.

Same fix applied to levelset's isovalue, which had the identical problem waiting.

How it was found

By testing BOSL2's actual delegation plan rather than assuming it would work. First the mechanism, against OpenSCAD 2026.02.01:

function norm(v) = 999;   // in the caller
echo(direct  = norm([3,4]));    // 999 -- the caller's own wins
echo(wrapped = _norm([3,4]));   // 5   -- the builtin, via a use<>d wrapper

Identical on both our engines. Then the full linear_solve pattern end to end, which is where the undef problem surfaced. It now runs clean:

square    = [1, 3]
det       = 5
overdet   = [-1.4, 3.1]                        (least squares)
underdet  = [0.333333, 2.33333, 2.66667]       (minimum norm)
singular  = []                                 (BOSL2's convention, preserved by their wrapper)

No warnings, both engines.

Two regression tests added, one per builtin. 1064 tests pass under both engines.

Version 0.50.0 → 0.50.1.

🤖 Generated with Claude Code